Gemma 4 26B-A4B · vLLM · NVFP4 + EAGLE3 · conc 2
status: done
Configuration
| Model | google/gemma-4-26B-A4B-it |
|---|---|
| Company | |
| Family | Gemma |
| Parameters | 26B / 4B (MoE) |
| Engine | vLLM + EAGLE3 (speculative decoding) |
| Quant / precision | NVFP4 |
| Why this quant | NVIDIA NVFP4 base (modelopt) + RedHatAI's official EAGLE3 speculator (RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3) — the EAGLE3 line of the crossover figure. |
| Download | nvidia/Gemma-4-26B-A4B-NVFP4 |
| Context window | 65536 |
| Input modalities | text, image (served text-only here) |
Measured results
| Prefill tok/s | 99.67 |
|---|---|
| Decode tok/s | 88.2 |
| Peak memory (GB) | 108.36 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head) |
| Completed | 2026-07-03 17:54 +0800 |
Full run command
VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 2 400 400 256 \
--speculative-config '{"model":"RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
# 152/400 prompts (hit the time cap), 0 errors, 402.2s. ready after 349s. TTFT median 146.8 ms, TPOT median 21.2 ms, req thr 0.378/s.
# SpecDecoding: mean acceptance length 2.12 (1.78-2.57) · avg draft acceptance 38% (26-52%) · per-position ~0.59/0.34/0.20.
Notes
Decode 88.2 tok/s aggregate at concurrency 2. NVFP4 + RedHatAI EAGLE3 speculator, conc 2, for the crossover figure.
- Image (pinned):
vllm/vllm-openai:nightly-aarch64@sha256:e414712fdc04…— the SINGLE image for all 24 cells of this figure. Ready after 349 s. - Workload: ShareGPT V3, concurrency 2. 152/400 completed, 0 errors before the 400 s time cap.
- Throughput: decode 88.2 tok/s aggregate, prefill 99.67 tok/s. TTFT median 146.8 ms, TPOT median 21.2 ms, req throughput 0.378/s.
- Spec-decode acceptance: mean acceptance length 2.12 (1.78-2.57) · avg draft acceptance 38% (26-52%) · per-position ~0.59/0.34/0.20 (num_speculative_tokens=3).
- Memory: 108.36 GB = vLLM
--gpu-memory-utilization 0.85reservation, not the model footprint.